Ronald Perry

1
kizamu

• Ronald Perry Sarah Frisken
• MERL

2
What is Kizamu?

• A computer-based sculpting system for creating
  digital characters
• Incorporates
• New algorithms
• Technical advances
• Novel interaction paradigms
• Focus
• High-end digital character design for the
  entertainment industry

3
Requirements for Digital Character Design

• Digital clay
• Clay-like intuitive to sculpt, represents both
  fine detail and organic shapes
• Digital can undo, script, duplicate, store
  permanently, etc.
• Responsive
• Interactive on standard hardware
• Fits into the animation production pipeline
• Accept scanned data as well as other standard
  representations
• Produce standard representations as output

4
Kizamu System

• Objects represented as Adaptively Sampled
  Distance Fields (ADFs)
• ADFs can be directly and intuitively edited
• ADFs represent fine detail and smooth organic
  surfaces
• ADFs support fast processing with a reasonable
  memory footprint
• Volumetric sculpting interface that exploits the
  distance field to provide intuitive interaction
• Accepts range data, triangle models, Bezier
  patches, and implicit functions as input
• Produces LOD triangle models as output

5
A Brief Overview of ADFs

• Distance fields
• A distance field is a scalar field that specifies
  the distance to the surface of a shape
• The distance may be signed to distinguish between
  the inside and outside of the shape

R shape
Distance field of R
6
Sampled Distance Fields
-130 -95 -62 -45 -31 -46 -57 -86
-129
-90
-90 -49 -2 17 25 16 -3
-43 -90
-71 -5 30 -4 -38 -32 -3
-46 12 1 -50 -93 -3
-65
20
2D shape with sampled distances to the surface
Regularly sampled distance values
2D distance field
7
Regularly Sampled Distance Fields

• Distance fields must be sampled at high enough
  rates to avoid aliasing (jagged edges)
• Very dense sampling is required when fine detail
  is present
• Regularly sampled distance fields require
  excessive memory when any fine detail is present

8
Adaptively Sampled Distance Fields (ADFs)

• Detail-directed sampling of a distance field
• High sampling rates only where needed
• Spatial data structure (e.g., an octree)
• Fast localization for efficient processing
• Reconstruction method (e.g., trilinear
  interpolation)
• For reconstructing the distance field and
  gradient from sampled distance values

9
Adaptively Sampled Distance Fields
A 2D crescent ADF and its quadtree data structure
10
Advantages of ADFs for Editing

• Represent both smooth surfaces and sharp corners
  without excessive memory
• Sculpting is direct, intuitive, and fast using
  simple Boolean operations
• Does not require control point manipulation or
  trimming
• The distance field can be used to enhance the
  user interface
• Guide the position and orientation of the
  sculpting tool
• Enable distance-based constraints for carving

11
What was Required to Build Kizamu

• Innovations in the sculpting interface
• Advances in ADF generation and ADF editing
• New approaches for ADF rendering
• Methods for generating ADFs directly from range
  data and converting ADFs to triangle models

12
Sculpting Interface

• Surface following
• Distance-based constraints
• Control-point editing

13
Sculpting Interface

• Surface following
• Distance-based constraints
• Control-point editing

14
Sculpting Interface

• Surface following
• Distance-based constraints
• Control-point editing

15
Bottom-up ADF Generation

• Requires too many distance computations, too much
  memory, pre-set resolution

Fully populate
Recursively coalesce
16
Top-down ADF Generation

• Requires expensive neighbor searches and
  redundant distance computation

Initialize root cell
Recursively subdivide
17
Tiled Generation

• Reduced memory requirements, better memory
  coherency and reduced computation
• Significantly faster (20x)
• 2 seconds vs 40 seconds for a level 9 ADF
• 7.6 seconds for a level 12 ADF
• More detail ((8x)3 higher resolution)
• level 12 and level 13 ADFs vs level 9 ADFs

18
Tiled Generation Method Outline

• Recursively subdivide root cell to a level L
• Cells at level L requiring further subdivision
  are appended to a list of candidate cells, C-list
  
• These candidate cells are recursively subdivided
  between levels L and 2L, where new candidate
  cells are produced and appended to C-list
• Repeat layered production of candidate cells (2L
  to 3L, etc.) until C-list is empty

19
Tiled Generation - Tiling

• For each candidate cell, computed and
  reconstructed distances are produced only as
  needed during subdivision
• These distances are stored in a tile, a regularly
  sampled volume
• Avoids recomputing distance values shared by
  neighboring cells. A corresponding volume of bit
  flags keeps track of valid distances in the tile.
• The tile resides in cache memory and its size
  determines L

20
Sculpting

• Sculpting is localized regeneration
• The ADF is regenerated inside cells that overlap
  the tools bounding volume
• Regeneration combines the distance fields of the
  ADF and the tool using CSG operations

21
Local Rendering

• Ray casting method
• Problems
• Too slow for local updates on low-end systems
• Woefully inadequate for global view changes

22
Adaptive Ray Casting

• Adaptive ray casting method
• Image divided into a hierarchy of tiles
• Tiles are subdivided according to perceptually
  based predicates
• Pixels in tiles greater than 1x1 are bilinearly
  interpolated
• Achieve 6x speedup (10x for supersampling)

23
Adaptive Ray Casting
Adaptively ray cast ADF
Rays cast to render part of the left image
24
Global Rendering of Point Models

• Method for generating point models from ADFs
• Seed each boundary leaf cell with randomly placed
  points, with number of points proportional to
  surface area within the cell
• Relax the points onto the ADF surface using the
  distance field and its gradient
• Optionally shade each point using the distance
  gradient
• Can generate 800,000 Phong illuminated points in
  0.2 seconds
• Point models are sufficient for positioning

25
Point-based Rendering
An ADF rendered as points at two different scales
26
Global Rendering using Triangles

• Convert ADFs to triangle models and render
  interactively with hardware
• Fast new triangulation algorithm creates triangle
  models from ADFs on-the-fly
• Watertight and oriented
• Good-quality triangles
• 200,000 triangles in 0.37 seconds
• Can create LOD models

27
Converting to Triangle Models

• Seed
• Assign a vertex to each boundary leaf cell of the
  ADF, initially placing vertices at cell centers
• Join
• Join vertices of neighboring cells to form
  triangles
• Relax
• Move vertices to the surface using the distance
  field
• Improve
• Move vertices over the surface towards their
  average neighbors' position to improve triangle
  quality

28
Creating LOD Triangle Models

• Adapt triangulation to generate LOD models
• Traverse octree from root to leaf cells
• Seed vertices in (possibly) non-leaf boundary
  cells that satisfy a minimum error criterion
• Ignore cells below these in the hierarchy

29
Converting Range Data to ADFs

• Capture 2D range images from multiple views

A range image
Range images collected from several view points
30
Converting Range Data to ADFs

• Combine distances from multiple range images to
  estimate ADF distance values

ADF generated from 14 synthetic range images
Combine distances from multiple range images
31
Results Detail
Boston road map sculpted with 2000 random
chisel strokes
32
Results Detail
ADF cow procedurally sculpted with 21,063 chisel
strokes
33
Results And More Detail
Flowers carved from a photograph using conversion
from range images to ADFs
34
Kizamu Demonstration
35
New Directions

• ADFs provide a single representation for both
  surface and volume models
• Extend Kizamu to support the creation of detailed
  volumetric digital characters

36
With Thanks

• Joe Letteri
• John Arnold
• Ray Jones
• Jackson Pope